Search

Graphics Processing Units (GPUs) exhibit significantly higher peak performance than conventional CPUs. However, in general only highly parallel algorithms can exploit their potential. In this scenario, the iterative solution to sparse linear systems of equations could be carried out quite efficiently on a GPU as it requires only matrix-by-vector products, dot products, and vector updates. However, to be really effective, any iterative solver needs to be properly preconditioned and this represents a major bottleneck for a successful GPU implementation. Due to its inherent parallelism, the factored sparse approximate inverse (FSAI) preconditioner represents an optimal candidate for the conjugate gradient-like solution of sparse linear systems. However, its GPU implementation requires a nontrivial recasting of multiple computational steps. We present our GPU version of the FSAI preconditioner along with a set of results that show how a noticeable speedup with respect to a highly tuned CPU counterpart is obtained. [ABSTRACT FROM AUTHOR]

In recent years the growing popularity of supercomputers has fostered the development of algorithms able to take advantage of the massive parallelism offered by multiple processors. Direct methods, though robust and computationally efficient, hardly exploit high degrees of parallelism. By contrast, Krylov methods preconditioned by Factored Sparse Approximate Inverses (FSAI) provide, at least in principle, a perfectly parallel approach but are often thwarted by an excessive set-up cost. In this paper we extend the concept of supernode from sparse LU factorizations to approximate inverses, and use it to accelerate the computation of an FSAI-type preconditioner. The numerical experiments on real-world problems show that the overall FSAI efficiency can be significantly increased while preserving its intrinsic parallelism. [ABSTRACT FROM AUTHOR]

Abstract: The Finite Element Method (FEM) is widely used in civil and mechanical engineering to simulate the behavior of complex structures and, more specifically, to predict stress and deformation fields of structural parts or mechanical bodies. In the former case, the coupling between different types of elements, such as beams, trusses, and shells, is often required, while in the latter fully 3D discretizations are typically used. For both, FEM leads to symmetric positive definite (SPD) matrices that, depending on the type of discretization and especially on the topology of the nodal connections, may be efficiently solved by either the Preconditioned Conjugate Gradient (PCG) or a direct solver such as the routine MA57 of the Harwell Software Library. Numerical experiments are shown and discussed where the effect of spatial discretization, different solution techniques, and a possible nodal reordering, is explored. The PCG preconditioner used is a variant of the incomplete Cholesky factorization with variable fill-in. It is shown that for structures with 1D or 2D connections, such as for example a bridge, MA57 performs usually better than PCG. In this case it is noted that some reorderings specifically designed and implemented for direct elimination methods can be very helpful for PCG as well as they yield a cheaper preconditioner and lead to a much faster PCG convergence. The main disadvantage is the need for an appropriate degree of fill-in for the preconditioner which turns out to be problem dependent and must be found empirically. However, in fully 3D problems, arising for example from the FE discretization of structural components or geomechanical structures, PCG outperforms MA57 while also requiring much less memory, and thus allowing for the use of much refined grids, if needed. With the aid of a large geomechanical problem it is shown that direct solvers may not be (even) used on serial computers due to their prohibitive computational cost with PCG the only viable alternative solver. [Copyright &y& Elsevier]

Abstract: Contact mechanics can be addressed numerically by finite elements using either a penalty formulation or the Lagrange multipliers. The penalty approach leads to a linearized symmetric positive definite system which can prove severely ill-conditioned, with the iterative solution to large 3D problems requiring expensive preconditioners to accelerate, or even to allow for, convergence. In the present paper a mixed constraint preconditioner (MCP) is developed and implemented for the iterative solution of contact problems based on the penalty approach. A theoretical bound for the eigenspectrum of the preconditioned matrix is provided depending on the user-specified parameters that control the preconditioner fill-in. The performance of MCP combined with the preconditioned conjugate gradient is studied in two real 3D test cases describing the geomechanics of faulted porous media and compared to a well-established preconditioner such as the incomplete Cholesky decomposition ILLT with optimal fill-in. Numerical results show that typically MCP outperforms ILLT especially in the most challenging problems where very large penalty parameters must be used. Finally, a modification aimed at simplifying the MCP implementation and increasing its robustness is developed and experimented with, showing that the modified MCP can be viewed as an efficient and viable alternative to both MCP and ILLT, especially in problems where the size of the contact surfaces is large with respect to the global domain. [Copyright &y& Elsevier]